[BugFix][K3] Skip moe_intermediate padding when EP is enabled (#51131)

ZeldaHuang · web-flow · commit beca88e59ea7 · 2026-08-05T09:22:55.000Z
diff --git a/vllm/models/kimi_k3/nvidia/model.py b/vllm/models/kimi_k3/nvidia/model.py
@@ -494,7 +494,7 @@ def __init__(
         min_moe_intermediate_per_partition = getattr(
             config, "min_moe_intermediate_per_partition", 256
         )
-        if self.tp_size > 1:
+        if self.tp_size > 1 and not vllm_config.parallel_config.enable_expert_parallel:
             moe_intermediate_per_partition = moe_intermediate_size // self.tp_size
             if moe_intermediate_per_partition < min_moe_intermediate_per_partition:
                 self.padded_moe_intermediate_size = (

Original file line number	Diff line number	Diff line change
`@@ -494,7 +494,7 @@ def __init__(`
`494`	`494`	`min_moe_intermediate_per_partition = getattr(`
`495`	`495`	`config, "min_moe_intermediate_per_partition", 256`
`496`	`496`	`)`
`497`		`- if self.tp_size > 1:`
	`497`	`+ if self.tp_size > 1 and not vllm_config.parallel_config.enable_expert_parallel:`
`498`	`498`	`moe_intermediate_per_partition = moe_intermediate_size // self.tp_size`
`499`	`499`	`if moe_intermediate_per_partition < min_moe_intermediate_per_partition:`
`500`	`500`	`self.padded_moe_intermediate_size = (`